Diagnosing and treating polyneuropathies often takes years, significantly impacting patients and increasing healthcare costs. Nerve sonography has improved diagnostics by 20% in some areas, but up to 30% of disease causes remain unidentified. This project aims to  increase the reliability of detecting treatable causes and optimize decision rules using robust and interpretable machine learning methods. 

ClinbrAIn PIs

Prof. Dr. Alexander Grimm

Dr. Natalie Winter

Prof. Dr. Matthias Bethge

Epilepsy is a chronic neurological disease characterised by a predisposition of the central nervous system to spontaneously develop epileptic seizures. The diagnosis of epilepsy relies on changes in behaviour during a seizure, and epilepsy-typical brain wave changes detectable in EEG. In clinical practice, behavioural assessment as well as diagnostic evaluation of the EEG has so far mostly been performed manually by clinical experts. This process, however, is not only labour and time intensive, but depends heavily on clinical expertise and is prone to errors. We focus on naturalistic neuroscience and aim to gain insights into the underlying mechanisms of epilepsy. In particular, we propose to investigate and utilize state-of-the-art AI based algorithms to extract meaning from behavioral measures, including text-based seizure descriptions - and EEG measures of epileptic patients.

ClinbrAIn Fellow

Meghal Dani

ClinbrAIn PIs

Prof. Dr. Zeynep Akata

Dr. Stephanie Liebe

Dr. Dr. Randolph Helfrich

This project aims to define spatial drug activity profiles in brain cancer by means of single-cell transcriptomic profiling. Common single-cell transcriptomics analyses typically evaluate expression profiles of single cells in isolation. To understand cellular dynamics in its tissue context it is necessary to complement the temporal and spatial context of these expression profiles.  However, this context is typically difficult to assess. One major reason for this was the lack of appropriate imaging methods which were only developed in recent years. Here we aim at developing machine learning methods to model the whole spatiotemporal process of brain tumors in response to different treatments, to gain insights into their spatially resolved effect in brain tissue, which in turn potentially motivates the development of new and more effective drugs or treatment regimens. 

ClinbrAIn Fellow

Marcello Zago

ClinbrAIn PIs

Prof. Dr. Ghazaleh Tabatabai

Prof. Dr. Manfred Claassen

Drug resistance in epilepsy patients poses a significant threat to the long-term treatment efficacy. Despite the existence of over 30 different anti-epileptic drugs with various targets and mechanisms of action, still, approximately one third of all patients with epilepsy remain drug resistant. We aim to identify clinical and genomic features of epilepsy patients that influence the response to anti-epileptic drugs with machine learning methods. Due to the scarcity of drug response data for less commonly administered drugs, we employ a Multi-Task learning framework to derive relevant information from other medications. 

ClinbrAIn Fellow

Julia Hellmig

ClinbrAIn PIs

Prof. Dr. Holger Lerche

Prof. Dr. Nico Pfeifer

Parkinson’s Disease presents itself highly heterogeneously in patients, where both rate of progression and phenotype outcomes are highly varying. We are developing a probabilistic machine learning model to elucidate progression patterns in Parkinson’s Disease patients. Longitudinal clinical data presents challenges due to missing data, noise, and varying measurement time intervals. We aim to develop a novel approach to dealing with these issues while prioritizing model interpretability. We are collaborating with clinical experts at Universitätsklinikum Tübingen to merge data from their study with the publicly available dataset from the Parkinson’s Progression Markers Initiative (PPMI). We then leverage this unique dataset to cluster patients according to rate of progression to varying phenotype endpoints. Modeling will focus on utilizing demographic features, well understand genetic markers, motor scores, biomarkers, cognitive assessments, and autonomic dysfunction assessments. 

ClinbrAIn Fellow

Sara Rajaram

ClinbrAIn PIs

Prof. Dr. Kathrin Brockmann

Prof. Dr. Fabian Sinz

Machine learning methods have demonstrated great potential for predicting disease progression over time, and an accurate prognosis can be highly relevant to both patients and physicians. However, most models do not provide faithful interpretations that allow humans to understand their predictions. To this end, we explore and develop deep learning models that are interpretable by design. In particular, we focus on the progression of age-related macular degeneration, an eye disease that can progress at different speeds for different groups of patients, eventually leading to loss of central vision. 

ClinbrAIn Fellow

Julius Gervelmeyer

ClinbrAIn PIs

Prof. Dr. Simon Clark

Prof. Dr. Philipp Berens

Publications

Gervelmeyer J, Müller S, Djoumessi K, Merle D, Clark SJ, Koch L, and Berens P. Interpretable-by-design Deep Survival Analysis for Disease Progression Modeling. medRxiv; 2024. DOI: 10.1101/2024.07.11.2431027.

Attention-Deficit/Hyperactivity Disorder (ADHD) is a complex neurodevelopmental condition affecting children and adolescents, and is often underdiagnosed despite its high prevalence. We aim to investigate the mechanisms of ADHD using data from the Adolescent Brain Cognitive Development (ABCD) study. This ten year longitudinal study tracks brain development and behaviour in over 11,500 children from ages 9 to 10 into young adulthood. We use Bayesian inference and Partially Observable Markov Decision Process (POMDP) models to analyze computational processes during cognitive tasks. We then use imaging data to find the neural correlates of the computational traits. By integrating theory—and data-driven methods, we aim to find individual differences in behaviour and neural systems across populations and stages of development. We hope this will ultimately provide crucial insights that may enable early diagnosis and the development of personalized interventions for affected individuals.

ClinbrAIn Fellow

Wenting Wang

ClinbrAIn PIs

Prof. Dr. Tobias Kaufmann

Prof. Dr. Peter Dayan

Transcranial magnetic stimulation combined with electroencephalography (TMS-EEG) is a promising tool for therapeutic brain stimulation. Aligning TMS with periods of increased or decreased cortical excitability can enhance its neuromodulatory effects, potentially improving therapeutic outcomes. However, current open-loop brain-state dependent methodologies suffer from high variability in stimulation responses. Moving beyond pre-defined biomarkers, we employ machine learning techniques to identify predictive brain activity patterns for immediate TMS effects on motor cortical networks. Ultimately, we aim to develop personalized TMS protocols for dynamically adjusting stimulation parameters based on real-time neural feedback to enhance the effectiveness of TMS interventions.

ClinbrAIn Fellow

Lisa Haxel

ClinbrAIn PIs

Prof. Dr. Ulf Ziemann

Prof. Dr. Jakob Macke

Publications

Haxel L, Belardinelli P, Ermolova M, et al. Decoding Motor Excitability in TMS using EEG-Features: An Exploratory Machine Learning Approach. bioRxiv; 2024. DOI: 10.1101/2024.02.27.582361. 

Deep Brain Stimulation (DBS) improves motor symptoms for Parkinson’s patients who do not sufficiently respond to medication or develop intolerance. The particular settings for DBS are decided based on standard heuristics, electrophysiological markers, and behavioral data from surgical procedures and ongoing treatment. Our goal is to enhance this treatment through advanced data analysis, optimization techniques, and the development of comprehensive neuronal models; altogether potentially improving the patients’ quality of life.  

Data Analysis: We aim to identify new electrophysiological biomarkers to supplement the existing markers, such as the commonly used excessive subthalamic beta oscillations. If the amount of data allows, we aim to use machine learning to define biomarkers automatically. 

Optimization: We aim to optimize various aspects of DBS treatment using our developed biomarkers. This includes better patient selection, precise electrode placement, and fine-tuning stimulation parameters such as electrode contacts, stimulation strength, patterns, and frequency. These optimizations will lead to more personalized and effective treatment plans for patients. 

Neuronal Models: Ultimately, our objective is to deepen our understanding of Parkinson’s disease by developing comprehensive neuronal models. These models will enable us to simulate and predict disease progression and treatment outcomes more accurately, paving the way for even more effective interventions. 

ClinbrAIn Fellow

Tim Schäfer

ClinbrAIn PIs

Prof. Dr. Alireza Gharabaghi

Prof. Dr. Anna Levina

Publications

Khajehabdollahi, S., Zeraati, R., Giannakakis, E., Schäfer, T. J., Martius, G., Levina, A. Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks. Twelfth International Conference on Learning Representations, ICLR 2024. Link.

In many neurological movement disorders (e.g. Parkinson's disease, cerebellar ataxia), gait disturbances are the first sign of the disease and affect daily life the most. For this reason, objective gait measurements are important patient-relevant performance indicators. Until now, gait analysis has been limited to laboratory examinations, which only reflects patients’ everyday life to a very limited extent, as movement behavior is examined in artificial situations. In order to establish ecologically valid performance markers evaluating treatment-responses in the patients’ everyday life, we develop multi-variate measures of ataxic gait using wearable sensors and machine learning. These gait measures demonstrate high sensitivity to small differences in disease severity in real-life walking, thus represent promising motor performance markers for natural history and treatment trials in ecologically valid contexts. The application of machine learning aims to increase clinical feasibility in future multi-center trials by implementing: 
(i)    the minimization of the number of necessary sensors 
(ii)   explainability and interpretability of the learnt classifier and  
(iii)  transfer learning to adapt the neural networks to analyze individual patients. 
This could lead to a system that, in the best case, only requires a smartphone or smartwatch, 
can be interpreted by clinicians, and enables individualized treatment for each patient. 

ClinbrAIn Fellow

Jens Seemann

ClinbrAIn PIs

Prof. Dr. Ludger Schöls

Prof. Dr. Martin Giese

Publications

L Beichert, J Seemann, …, R Schüle (2024): Towards patient-relevant, trial-ready digital motor outcomes for SPG7: a cross-sectional prospective multi-center study (PROSPAX). medRrxiv, 2024. DOI: 10.1101/2024.01.09.24301064.
 
J Seemann, L Daghsen, M Cazier, JC Lamy, ML Welter, MA Giese, M Synofzik, A Durr, W Ilg, G Coarelli (2024): Digital gait measures capture 1-year progression in early-stage spinocerebellar ataxia type 2. Movement Disorders 39 (5), 788-797.  DOI: 10.1002/mds.29757.

 
J Seemann, T Loris, L Weber, M Synofzik, MA Giese, W Ilg (2023): One Hip Wonder: 1D-CNNs Reduce Sensor Requirements for Everyday Gait Analysis. International Conference on Artificial Neural Networks, 346-357.  DOI: 10.1007/978-3-031-44204-9_29.
 
J Seemann, A Traschütz, W Ilg, M Synofzik (2023): 4-Aminopyridine improves real-life gait performance in SCA27B on a single-subject level: a prospective n-of-1 treatment experience. Journal of Neurology, 1-6.  DOI: 10.1007/s00415-023-11868-y.